@maintainabilityai/research-runner 0.1.7 → 0.1.9

package/dist/llm/anthropic-client.js

@@ -23,8 +23,12 @@ async function callAnthropic(opts) {
         throw new Error('ANTHROPIC_API_KEY missing — set the env var or pass apiKey directly');
     }
     const fetchImpl = opts.fetchImpl ?? globalThis.fetch;
-    // Match github-models default — 8K-token synth responses can run 60–90s.
-    const timeoutMs = opts.timeoutMs ?? 120_000;
+    // Sonnet at ~50–80 tok/s × 8K output = 100–160s, plus prompt-processing.
+    // 120s aborted real synth runs mid-stream; 240s gives headroom for
+    // slow days. (Real fix is streaming so the connection stays alive,
+    // but the cost/benefit isn't worth it yet — single-shot is fine
+    // until we hit 240s legitimately.)
+    const timeoutMs = opts.timeoutMs ?? 240_000;
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(), timeoutMs);
     let response;

package/dist/llm/github-models-client.js

@@ -25,10 +25,13 @@ async function callGitHubModels(opts) {
     }
     const fetchImpl = opts.fetchImpl ?? globalThis.fetch;
     const endpoint = opts.endpoint ?? DEFAULT_ENDPOINT;
-    // Synthesis prompts can produce 8K-token responses (and the "custom"
-    // tier models like gpt-5-mini can return much more), which routinely
-    // take 60–120s. Default to 120s so we don't abort mid-stream.
-    const timeoutMs = opts.timeoutMs ?? 120_000;
+    // Synthesis prompts can produce 8K+ output tokens (and "custom"-tier
+    // models like gpt-5-chat can return much more), which routinely take
+    // 100–180s. Match the Anthropic client at 240s for headroom on slow
+    // days. (Real fix is streaming so the connection stays alive, but
+    // the cost/benefit isn't worth it yet — single-shot is fine until
+    // we hit 240s legitimately.)
+    const timeoutMs = opts.timeoutMs ?? 240_000;
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(), timeoutMs);
     const messages = [];

package/dist/llm/llm-router.js

@@ -16,62 +16,96 @@ const MODEL_BY_TIER = {
     // Models access through a Copilot subscription (GMT path).
     synth: { anthropic: 'claude-sonnet-4-6', githubModels: 'openai/gpt-5-chat' },
 };
+/**
+ * Write progress to stderr so the GitHub Actions job shows the
+ * "tried X, falling back to Y" story when the primary provider fails.
+ * Silenced when RESEARCH_RUNNER_QUIET=1 (tests).
+ */
+function progress(msg) {
+    if (process.env.RESEARCH_RUNNER_QUIET === '1') {
+        return;
+    }
+    const ts = new Date().toISOString().slice(11, 19);
+    process.stderr.write(`[research-runner ${ts}] ${msg}\n`);
+}
+async function callAnthropicTier(opts, tierModels) {
+    if (!opts.anthropicApiKey) {
+        throw new Error(`callLlm: provider=anthropic requires anthropicApiKey (set ANTHROPIC_API_KEY).`);
+    }
+    const r = await (0, anthropic_client_1.callAnthropic)({
+        apiKey: opts.anthropicApiKey,
+        model: tierModels.anthropic,
+        system: opts.system,
+        prompt: opts.prompt,
+        maxTokens: opts.maxTokens,
+        temperature: opts.temperature,
+        fetchImpl: opts.fetchImpl,
+    });
+    return {
+        provider: 'anthropic',
+        model: tierModels.anthropic,
+        text: r.text,
+        inputTokens: r.inputTokens,
+        outputTokens: r.outputTokens,
+        costUsd: r.costUsd,
+        httpStatus: r.httpStatus,
+    };
+}
+async function callGitHubModelsTier(opts, tierModels) {
+    if (!opts.githubToken) {
+        throw new Error(`callLlm: provider=github-models requires githubToken (set GITHUB_TOKEN; workflow needs \`permissions: models: read\`).`);
+    }
+    const r = await (0, github_models_client_1.callGitHubModels)({
+        token: opts.githubToken,
+        model: tierModels.githubModels,
+        system: opts.system,
+        prompt: opts.prompt,
+        maxTokens: opts.maxTokens,
+        temperature: opts.temperature,
+        fetchImpl: opts.fetchImpl,
+    });
+    return {
+        provider: 'github-models',
+        model: tierModels.githubModels,
+        text: r.text,
+        inputTokens: r.inputTokens,
+        outputTokens: r.outputTokens,
+        costUsd: r.costUsd,
+        httpStatus: r.httpStatus,
+    };
+}
 async function callLlm(opts) {
     const tierModels = MODEL_BY_TIER[opts.tier];
-    // Hybrid routing: GitHub Models free tier caps requests at ~8K input
-    // tokens — too small for the synthesis step (full brief + every search
-    // result + mesh context routinely exceeds that). When the brief asks
-    // for github-models AND an Anthropic key is available, route synth →
-    // Anthropic and keep plan (small prompt) on github-models. Caller can
-    // force pure github-models by not setting anthropicApiKey.
-    const effectiveProvider = opts.provider === 'github-models' && opts.tier === 'synth' && opts.anthropicApiKey
-        ? 'anthropic'
-        : opts.provider;
-    if (effectiveProvider === 'anthropic') {
-        if (!opts.anthropicApiKey) {
-            throw new Error(`callLlm: provider=anthropic requires anthropicApiKey (set ANTHROPIC_API_KEY).`);
+    // Try-then-fall-back routing for github-models.
+    //
+    // When the brief asks for github-models AND an Anthropic key is also
+    // available, we try GitHub Models FIRST (respects the user's stated
+    // provider preference — they're often paying for Copilot Pro). If it
+    // fails for ANY reason (8K cap 413, custom-tier 403, timeout, 5xx,
+    // network), fall back to Anthropic. The synth tier is where this
+    // matters most — plan-tier prompts are small and almost never fail.
+    //
+    // Why fall back on every error vs. only on 413: we'd rather have a
+    // working synth via Anthropic than a half-failed run. The progress
+    // log makes the "tried X, fell back to Y" story explicit so it's
+    // easy to debug.
+    //
+    // anthropic provider is pass-through (no fallback target).
+    if (opts.provider === 'github-models') {
+        try {
+            return await callGitHubModelsTier(opts, tierModels);
         }
-        const r = await (0, anthropic_client_1.callAnthropic)({
-            apiKey: opts.anthropicApiKey,
-            model: tierModels.anthropic,
-            system: opts.system,
-            prompt: opts.prompt,
-            maxTokens: opts.maxTokens,
-            temperature: opts.temperature,
-            fetchImpl: opts.fetchImpl,
-        });
-        return {
-            provider: 'anthropic',
-            model: tierModels.anthropic,
-            text: r.text,
-            inputTokens: r.inputTokens,
-            outputTokens: r.outputTokens,
-            costUsd: r.costUsd,
-            httpStatus: r.httpStatus,
-        };
-    }
-    if (effectiveProvider === 'github-models') {
-        if (!opts.githubToken) {
-            throw new Error(`callLlm: provider=github-models requires githubToken (set GITHUB_TOKEN; workflow needs \`permissions: models: read\`).`);
+        catch (err) {
+            if (!opts.anthropicApiKey) {
+                throw err; // no fallback available
+            }
+            const msg = err instanceof Error ? err.message : String(err);
+            progress(`⚠ github-models failed on tier=${opts.tier} — falling back to Anthropic. Cause: ${msg.slice(0, 200)}`);
+            return await callAnthropicTier(opts, tierModels);
         }
-        const r = await (0, github_models_client_1.callGitHubModels)({
-            token: opts.githubToken,
-            model: tierModels.githubModels,
-            system: opts.system,
-            prompt: opts.prompt,
-            maxTokens: opts.maxTokens,
-            temperature: opts.temperature,
-            fetchImpl: opts.fetchImpl,
-        });
-        return {
-            provider: 'github-models',
-            model: tierModels.githubModels,
-            text: r.text,
-            inputTokens: r.inputTokens,
-            outputTokens: r.outputTokens,
-            costUsd: r.costUsd,
-            httpStatus: r.httpStatus,
-        };
     }
-    throw new Error(`callLlm: provider "${effectiveProvider}" not yet implemented (phase 2c.1 ships anthropic + github-models; openai + azure-openai land later).`);
+    if (opts.provider === 'anthropic') {
+        return await callAnthropicTier(opts, tierModels);
+    }
+    throw new Error(`callLlm: provider "${opts.provider}" not yet implemented (phase 2c.1 ships anthropic + github-models; openai + azure-openai land later).`);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maintainabilityai/research-runner",
-  "version": "0.1.7",
+  "version": "0.1.9",
   "description": "Research + PRD agent runner — orchestrates the Archeologist and PRD pipelines for the MaintainabilityAI governance mesh",
   "license": "MIT",
   "author": "MaintainabilityAI",